70 return pNetworkImpl->AddComparisonLayer(comparisonDescriptor, name);
77 return pNetworkImpl->AddConcatLayer(concatDescriptor, name);
86 return pNetworkImpl->AddConvolution2dLayer(convolution2dDescriptor, weights, biases, name);
95 return pNetworkImpl->AddConvolution2dLayer(convolution2dDescriptor, weights, biases, name);
105 return pNetworkImpl->AddConvolution2dLayer(convolution2dDescriptor,
115 return pNetworkImpl->AddDepthToSpaceLayer(depthToSpaceDescriptor, name);
125 return pNetworkImpl->AddDepthwiseConvolution2dLayer(convolution2dDescriptor, weights, biases, name);
135 return pNetworkImpl->AddDepthwiseConvolution2dLayer(convolution2dDescriptor, weights, biases, name);
145 return pNetworkImpl->AddDepthwiseConvolution2dLayer(convolution2dDescriptor, weights,
161 return pNetworkImpl->AddDetectionPostProcessLayer(descriptor, anchors, name);
168 return pNetworkImpl->AddElementwiseUnaryLayer(elementwiseUnaryDescriptor, name);
175 return pNetworkImpl->AddFillLayer(fillDescriptor, name);
183 return pNetworkImpl->AddFullyConnectedLayer(fullyConnectedDescriptor,
194 return pNetworkImpl->AddFullyConnectedLayer(fullyConnectedDescriptor,
205 return pNetworkImpl->AddFullyConnectedLayer(fullyConnectedDescriptor,
216 return pNetworkImpl->AddFullyConnectedLayer(fullyConnectedDescriptor, weights, biases, name);
222 return pNetworkImpl->AddPermuteLayer(permuteDescriptor, name);
228 return pNetworkImpl->AddBatchToSpaceNdLayer(batchToSpaceNdDescriptor, name);
234 return pNetworkImpl->AddPooling2dLayer(pooling2dDescriptor, name);
240 return pNetworkImpl->AddActivationLayer(activationDescriptor, name);
246 return pNetworkImpl->AddNormalizationLayer(normalizationDescriptor, name);
251 return pNetworkImpl->AddSliceLayer(sliceDescriptor, name);
256 return pNetworkImpl->AddSoftmaxLayer(softmaxDescriptor, name);
262 return pNetworkImpl->AddSplitterLayer(splitterDescriptor, name);
273 return pNetworkImpl->AddConcatLayer(mergerDescriptor, name);
298 return pNetworkImpl->AddBatchNormalizationLayer(desc, mean, variance, beta, gamma, name);
317 return pNetworkImpl->AddResizeLayer(resizeDescriptor, name);
323 return pNetworkImpl->AddResizeLayer(resizeDescriptor, name);
329 return pNetworkImpl->AddReduceLayer(reduceDescriptor, name);
335 return pNetworkImpl->AddInstanceNormalizationLayer(desc, name);
341 return pNetworkImpl->AddL2NormalizationLayer(desc, name);
347 return pNetworkImpl->AddLogSoftmaxLayer(logSoftmaxDescriptor, name);
359 return pNetworkImpl->AddReshapeLayer(reshapeDescriptor, name);
365 return pNetworkImpl->AddSpaceToBatchNdLayer(spaceToBatchNdDescriptor, name);
371 return pNetworkImpl->AddSpaceToDepthLayer(spaceToDepthDescriptor, name);
387 return pNetworkImpl->AddLstmLayer(descriptor, params, name);
407 return pNetworkImpl->AddMeanLayer(meanDescriptor, name);
424 return pNetworkImpl->AddStridedSliceLayer(stridedSliceDescriptor, name);
450 return pNetworkImpl->AddGatherLayer(gatherDescriptor, name);
474 return pNetworkImpl->AddTransposeConvolution2dLayer(descriptor, weights, biases, name);
480 return pNetworkImpl->AddTransposeLayer(transposeDescriptor, name);
498 return pNetworkImpl->AddQuantizedLstmLayer(params, name);
505 return pNetworkImpl->AddQLstmLayer(descriptor, params, name);
511 return pNetworkImpl->AddLogicalBinaryLayer(descriptor, name);
526 return new INetwork(networkOptions);
579 return m_Graph->SerializeToDot(stream);
583 Optional<std::vector<std::string>&> errorMessages)
585 std::stringstream fullErrorMessage;
586 fullErrorMessage <<
"ERROR: " << errorMessage;
590 errorMessages.value().push_back(fullErrorMessage.str());
595 Optional<std::vector<std::string>&> warningMessages)
597 std::stringstream fullWarningMessage;
598 fullWarningMessage <<
"WARNING: " << warningMessage;
602 warningMessages.value().push_back(fullWarningMessage.str());
609 Optional<std::vector<std::string>&> errMessages)
611 std::stringstream failureMsg;
623 bool noErrors =
true;
625 for (
unsigned int i = 0; i < numOutputs; i++) {
631 std::stringstream ss;
633 <<
" (" << layer->
GetNameStr() <<
") is of type" 634 <<
" Quantized 8 bit but its scale parameter has not been set";
642 std::stringstream ss;
643 ss <<
"Quantization parameters for Softmax layer (Scale: " <<
645 ") are incorrect and have been updated to Scale: 0.00390625 and Offset: 0";
656 template <
typename LayerT>
659 LayerT* layer = PolymorphicDowncast<LayerT*>(l);
670 layer->m_Weight->template GetConstTensor<armnn::BFloat16>(), info.
GetNumElements(), newValues.data());
686 const std::vector<BackendId>& availablePreferredBackends,
687 std::string& reasonIfUnsupported,
688 Optional<std::vector<std::string>&> errMessages)
693 auto ReturnError = [&](
const Layer* layer)
710 std::vector<ConvertFp16ToFp32Layer*> convertFp16ToFp32Layers;
713 convertFp16ToFp32Layers =
718 std::vector<ConvertFp32ToFp16Layer*> convertFp32ToFp16Layers;
721 convertFp32ToFp16Layers =
726 auto AssignFirstSupportedBackend = [&](
Layer* layer,
BackendId preferredBackend)
728 bool supportedBackendFound =
false;
729 std::string reasonIfUnsupported;
735 reasonIfUnsupported))
737 supportedBackendFound =
true;
741 for (
const auto& backend : availablePreferredBackends)
744 if (backend == preferredBackend)
752 reasonIfUnsupported))
754 supportedBackendFound =
true;
760 return supportedBackendFound;
765 if (!AssignFirstSupportedBackend(convertLayer, backend))
767 return ReturnError(convertLayer);
773 if (!AssignFirstSupportedBackend(convertLayer, backend))
775 return ReturnError(convertLayer);
789 std::vector<ConvertBf16ToFp32Layer*> convertBf16ToFp32Layers;
792 convertBf16ToFp32Layers =
796 ConvertBf16ToFp32Weight<Convolution2dLayer>(layer);
800 ConvertBf16ToFp32Weight<FullyConnectedLayer>(layer);
805 std::vector<ConvertFp32ToBf16Layer*> convertFp32ToBf16Layers;
808 convertFp32ToBf16Layers =
813 auto AssignFirstSupportedBackend = [&](
Layer* layer,
BackendId preferredBackend)
815 bool supportedBackendFound =
false;
816 std::string reasonIfUnsupported;
822 reasonIfUnsupported))
824 supportedBackendFound =
true;
828 for (
const auto& backend : availablePreferredBackends)
831 if (backend == preferredBackend)
839 reasonIfUnsupported))
841 supportedBackendFound =
true;
847 return supportedBackendFound;
852 if (!AssignFirstSupportedBackend(convertLayer, backend))
854 return ReturnError(convertLayer);
860 if (!AssignFirstSupportedBackend(convertLayer, backend))
862 return ReturnError(convertLayer);
870 std::stringstream warningMsg;
872 <<
" is not supported on requested backend " << layer->
GetBackendId().
Get()
875 <<
" (reason: " << reasonIfUnsupported
876 <<
"), falling back to the next backend.";
892 Optional<std::vector<std::string>&> errMessages)
897 auto ReturnError = [&](
const Layer* layer)
904 if (availablePreferredBackends.empty())
906 std::stringstream failureMsg;
907 failureMsg <<
"No preferred backends are available";
914 for (
auto it = firstLayer; it != lastLayer; ++it)
919 layer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo().GetDataType();
921 layer->GetOutputSlot(0).GetTensorInfo().GetDataType();
923 std::string reasonIfUnsupported;
933 if (layer->GetBackendHint().has_value() &&
938 layer->GetBackendHint().value(),
941 availablePreferredBackends,
951 for (
const auto& backend : availablePreferredBackends)
953 if (layer->GetBackendHint().has_value() &&
954 layer->GetBackendHint().value() == backend)
965 availablePreferredBackends,
1001 layer->SetBackendId(cpuBackendId);
1006 return ReturnError(layer);
1017 Optional<std::vector<std::string>&> errMessages)
1035 auto backendFactory = backendRegistry.GetFactory(selectedBackend);
1036 auto backendObjPtr = backendFactory();
1039 backendObjPtr->RegisterTensorHandleFactories(handleFactoryRegistry);
1041 backends[backendObjPtr->GetId()] = std::move(backendObjPtr);
1051 Optional<std::vector<std::string>&> errMessages)
1063 auto backendObjPtr = backends.find(selectedBackend)->second.get();
1070 [&backendObjPtr](
const Layer& layer)
1076 if (subgraphs.empty())
1083 for (
auto& subgraph : subgraphs)
1086 OptimizationViews optimizationViews = backendObjPtr->OptimizeSubgraphView(*subgraph, modelOptions);
1093 SubgraphView& replacementSubgraph = substitution.m_ReplacementSubgraph;
1094 SubgraphView& substitutableSubgraph = substitution.m_SubstitutableSubgraph;
1098 std::for_each(replacementSubgraph.
begin(), replacementSubgraph.
end(), [&selectedBackend](
Layer* l)
1101 l->SetBackendId(selectedBackend);
1107 std::stringstream warningMsg;
1108 warningMsg <<
"Some sub-graph(s) failed to optimized on " << backendObjPtr->GetId() <<
" backend.";
1113 if (!backendObjPtr->GetId().IsCpuRef())
1116 settingsCopy.m_IgnoredBackends.insert(backendObjPtr->GetId());
1123 std::stringstream subgraphMsg;
1124 subgraphMsg <<
"Re-assigning backends to " << failedSubgraph.GetLayers().size()
1125 <<
" layers inside sub-graph " << count++;
1132 if (reassignmentResult.m_Error)
1155 if (srcFactory && dstFactory &&
1181 if (frmBackend == backends.end() ||
1182 !frmBackend->second->SupportsTensorAllocatorAPI())
1189 std::map<ITensorHandleFactory::FactoryId, int> factoryScores;
1196 const Layer& connectedLayer = connection->GetOwningLayer();
1198 auto toBackend = backends.find(connectedLayer.
GetBackendId());
1199 ARMNN_ASSERT_MSG(toBackend != backends.end(),
"Backend id not found for the connected layer");
1201 if (!toBackend->second.get()->SupportsTensorAllocatorAPI())
1207 auto dstPrefs = toBackend->second.get()->GetHandleFactoryPreferences();
1208 for (
auto&& dst : dstPrefs)
1222 auto it = factoryScores.find(dst);
1223 if (it == factoryScores.end())
1226 factoryScores[dst] = 0;
1235 factoryScores[dst]++;
1238 if (factoryScores[dst] > topScore)
1240 topScore = factoryScores[dst];
1269 if (frmBackend == backends.end() ||
1270 !frmBackend->second->SupportsTensorAllocatorAPI())
1275 bool outputConnection =
false;
1278 const Layer& connectedLayer = connection->GetOwningLayer();
1281 outputConnection =
true;
1289 std::map<ITensorHandleFactory::FactoryId, int> factoryScores;
1290 for (
auto&& pref : srcPrefs)
1295 if (outputConnection)
1298 bool fallbackConnection =
false;
1301 if (inputSlot.GetConnectedOutputSlot()->GetOwningLayer().GetBackendId() != layer.
GetBackendId())
1303 fallbackConnection =
true;
1306 if (fallbackConnection)
1310 if (!factoryCap.empty())
1320 if (!outputConnection)
1324 if (!factoryCap.empty())
1343 auto it = factoryScores.find(pref);
1344 if (it == factoryScores.end())
1347 factoryScores[pref] = 0;
1354 const Layer& connectedLayer = connection->GetOwningLayer();
1356 auto toBackend = backends.find(connectedLayer.
GetBackendId());
1357 ARMNN_ASSERT_MSG(toBackend != backends.end(),
"Backend id not found for the connected layer");
1359 auto dstPrefs = toBackend->second.get()->GetHandleFactoryPreferences();
1360 for (
auto&& src : srcPrefs)
1362 if (factoryScores.find(src) == factoryScores.end())
1367 for (
auto&& dst : dstPrefs)
1372 factoryScores[src]++;
1380 int minScore = std::numeric_limits<int>::max();
1381 for (
auto it : factoryScores)
1383 minScore = std::min(minScore, it.second);
1387 std::vector<ITensorHandleFactory::FactoryId> optimalFactories;
1388 for (
auto it : factoryScores)
1390 if (it.second == minScore)
1392 optimalFactories.push_back(it.first);
1397 for (
auto&& srcPref : srcPrefs)
1399 for (
auto&& comp : optimalFactories)
1401 if (comp == srcPref)
1414 const Layer& connectedLayer,
1418 auto toBackend = backends.find(connectedLayer.
GetBackendId());
1419 ARMNN_ASSERT_MSG(toBackend != backends.end(),
"Backend id not found for the connected layer");
1421 auto dstPrefs = toBackend->second.get()->GetHandleFactoryPreferences();
1444 for (
auto&& pref : dstPrefs)
1446 if (pref == srcFactoryId)
1456 for (
auto&& pref : dstPrefs)
1475 if (srcCapability.empty() && dstCapability.empty() && srcFallback.empty() && dstFallback.empty())
1486 for (
auto&& pref : dstPrefs)
1504 Optional<std::vector<std::string>&> errMessages)
1508 optGraph.
ForEachLayer([&backends, ®istry, &result, &errMessages, importEnabled](
Layer* layer)
1539 unsigned int connectionIdx = 0;
1542 const Layer& connectedLayer = connection->GetOwningLayer();
1545 registry, importEnabled);
1552 errMessages.value().emplace_back(
"Could not find valid strategy required for compatibility" 1553 " between backends.");
1569 const std::vector<BackendId>& backendPreferences,
1572 Optional<std::vector<std::string>&> messages)
1574 if (backendPreferences.empty())
1584 std::unique_ptr<Graph> graph = std::make_unique<Graph>(inNetwork.
pNetworkImpl->GetGraph());
1595 using namespace optimizations;
1640 std::stringstream failureMsg;
1641 failureMsg <<
"None of the preferred backends " << backendPreferences
1659 if (assignBackendsResult.
m_Error)
1674 if (backendOptimizationResult.
m_Error)
1690 tensorHandleFactoryRegistry,
1710 auto backendPtr = factoryFun();
1714 auto backendSpecificOptimizations = backendPtr->GetOptimizations();
1717 if (!backendSpecificOptimizations.empty())
1725 bool NetworkImpl::GetShapeInferenceMethod()
1727 if (m_NetworkOptions.size() > 0 && m_NetworkOptions[0].GetBackendId().Get() ==
"ShapeInferenceMethod")
1729 return m_NetworkOptions[0].GetOption(0).GetValue().AsBool();
1735 : m_NetworkOptions(networkOptions),
1736 m_Graph(
std::make_unique<
Graph>(GetShapeInferenceMethod()))
1751 return m_Graph->AddLayer<
InputLayer>(id, name);
1762 return m_Graph->AddLayer<
CastLayer>(name);
1768 return m_Graph->AddLayer<
ComparisonLayer>(comparisonDescriptor, name);
1780 return m_Graph->AddLayer<
FillLayer>(fillDescriptor, name);
1798 const auto layer = m_Graph->AddLayer<
FullyConnectedLayer>(fullyConnectedDescriptor, name);
1800 if (fullyConnectedDescriptor.m_ConstantWeights)
1802 layer->
m_Weight = std::make_shared<ScopedTensorHandle>(weights.
value());
1803 if (fullyConnectedDescriptor.m_BiasEnabled)
1805 layer->m_Bias = std::make_shared<ScopedTensorHandle>(biases.
value());
1817 return AddFullyConnectedLayerImpl(fullyConnectedDescriptor, weights, biases, name);
1826 return AddFullyConnectedLayerImpl(fullyConnectedDescriptor, optionalWeights, biases, name);
1835 return AddFullyConnectedLayerImpl(fullyConnectedDescriptor, optionalWeights, biases, name);
1845 return AddFullyConnectedLayerImpl(fullyConnectedDescriptor, optionalWeights, optionalBiases, name);
1851 return m_Graph->AddLayer<
ConcatLayer>(concatDescriptor, name);
1864 const auto layer = m_Graph->AddLayer<
Convolution2dLayer>(convolution2dDescriptor, name);
1866 layer->
m_Weight = std::make_shared<ScopedTensorHandle>(weights);
1868 if (convolution2dDescriptor.m_BiasEnabled)
1870 layer->m_Bias = std::make_shared<ScopedTensorHandle>(biases.
value());
1881 return AddConvolution2dLayerImpl(convolution2dDescriptor, weights, biases, name);
1889 return AddConvolution2dLayerImpl(convolution2dDescriptor, weights, biases, name);
1898 return AddConvolution2dLayerImpl(convolution2dDescriptor, weights, optionalBiases, name);
1914 layer->
m_Weight = std::make_shared<ScopedTensorHandle>(weights);
1916 if (convolution2dDescriptor.m_BiasEnabled)
1918 layer->m_Bias = std::make_shared<ScopedTensorHandle>(biases.
value());
1936 return AddDepthwiseConvolution2dLayerImpl(convolution2dDescriptor, weights, biases, name);
1945 return AddDepthwiseConvolution2dLayerImpl(convolution2dDescriptor, weights, biases, name);
1955 return AddDepthwiseConvolution2dLayerImpl(convolution2dDescriptor, weights, optionalBiases, name);
1971 return m_Graph->AddLayer<
PermuteLayer>(permuteDescriptor, name);
1977 return m_Graph->AddLayer<
Pooling2dLayer>(pooling2dDescriptor, name);
1983 return m_Graph->AddLayer<
ActivationLayer>(activationDescriptor, name);
1989 return m_Graph->AddLayer<
ArgMinMaxLayer>(argMinMaxDescriptor, name);
1993 normalizationDescriptor,
2001 return m_Graph->AddLayer<
SliceLayer>(sliceDescriptor, name);
2007 return m_Graph->AddLayer<
SoftmaxLayer>(softmaxDescriptor, name);
2013 return m_Graph->AddLayer<
SplitterLayer>(splitterDescriptor, name);
2061 layer->
m_Mean = std::make_shared<ScopedTensorHandle>(mean);
2062 layer->m_Variance = std::make_shared<ScopedTensorHandle>(variance);
2063 layer->m_Beta = std::make_shared<ScopedTensorHandle>(beta);
2064 layer->m_Gamma = std::make_shared<ScopedTensorHandle>(gamma);
2071 return m_Graph->AddLayer<
RankLayer>(name);
2077 return m_Graph->AddLayer<
ReduceLayer>(reduceDescriptor, name);
2091 return m_Graph->AddLayer<
ResizeLayer>(resizeDescriptor, name);
2096 return m_Graph->AddLayer<
ResizeLayer>(resizeDescriptor, name);
2121 layer->
m_LayerOutput = std::make_shared<ScopedTensorHandle>(input);
2129 return m_Graph->AddLayer<
ReshapeLayer>(reshapeDescriptor, name);
2153 const auto layer = m_Graph->AddLayer<
LstmLayer>(descriptor, name);
2158 layer->m_BasicParameters.m_InputToCellWeights =
2160 layer->m_BasicParameters.m_InputToOutputWeights =
2162 layer->m_BasicParameters.m_RecurrentToForgetWeights =
2164 layer->m_BasicParameters.m_RecurrentToCellWeights =
2166 layer->m_BasicParameters.m_RecurrentToOutputWeights =
2168 layer->m_BasicParameters.m_ForgetGateBias =
2170 layer->m_BasicParameters.m_CellBias =
2171 std::make_shared<ScopedTensorHandle>(*(params.
m_CellBias));
2172 layer->m_BasicParameters.m_OutputGateBias =
2176 if(!descriptor.m_CifgEnabled)
2181 "when CIFG is disabled.");
2186 "AddLstmLayer: Recurrent To Input Weights cannot be NULL " 2187 "when CIFG is disabled.");
2192 "when CIFG is disabled.");
2194 layer->m_CifgParameters.m_InputToInputWeights =
2196 layer->m_CifgParameters.m_RecurrentToInputWeights =
2198 layer->m_CifgParameters.m_InputGateBias =
2203 if(descriptor.m_ProjectionEnabled)
2208 "when projection is enabled.");
2210 layer->m_ProjectionParameters.m_ProjectionWeights =
2214 layer->m_ProjectionParameters.m_ProjectionBias =
2220 if(descriptor.m_PeepholeEnabled)
2222 if(!descriptor.m_CifgEnabled)
2227 "when Peephole is enabled and CIFG disabled.");
2230 layer->m_PeepholeParameters.m_CellToInputWeights =
2237 "when Peephole is enabled.");
2242 "when Peephole is enabled.");
2245 layer->m_PeepholeParameters.m_CellToForgetWeights =
2247 layer->m_PeepholeParameters.m_CellToOutputWeights =
2252 if(descriptor.m_LayerNormEnabled)
2254 if(!descriptor.m_CifgEnabled)
2259 "when layer normalization is enabled and CIFG disabled.");
2261 layer->m_LayerNormParameters.m_InputLayerNormWeights =
2268 "when layer normalization is enabled.");
2273 "when layer normalization is enabled.");
2278 "when layer normalization is enabled.");
2280 layer->m_LayerNormParameters.m_ForgetLayerNormWeights =
2282 layer->m_LayerNormParameters.m_CellLayerNormWeights =
2284 layer->m_LayerNormParameters.m_OutputLayerNormWeights =
2302 return m_Graph->AddLayer<
MeanLayer>(meanDescriptor,name);
2307 return m_Graph->AddLayer<
PadLayer>(padDescriptor,name);
2350 return m_Graph->AddLayer<
GatherLayer>(gatherDescriptor, name);
2380 layer->
m_Weight = std::make_shared<ScopedTensorHandle>(weights);
2382 if (descriptor.m_BiasEnabled)
2384 layer->m_Bias = std::make_shared<ScopedTensorHandle>(biases.
value());
2393 return m_Graph->AddLayer<
TransposeLayer>(transposeDescriptor, name);
2399 return m_Graph->AddLayer<
StackLayer>(stackDescriptor, name);
2417 layer->m_QuantizedLstmParameters.m_InputToForgetWeights =
2419 layer->m_QuantizedLstmParameters.m_InputToCellWeights =
2421 layer->m_QuantizedLstmParameters.m_InputToOutputWeights =
2425 layer->m_QuantizedLstmParameters.m_RecurrentToInputWeights =
2427 layer->m_QuantizedLstmParameters.m_RecurrentToForgetWeights =
2429 layer->m_QuantizedLstmParameters.m_RecurrentToCellWeights =
2431 layer->m_QuantizedLstmParameters.m_RecurrentToOutputWeights =
2435 layer->m_QuantizedLstmParameters.m_InputGateBias =
2437 layer->m_QuantizedLstmParameters.m_ForgetGateBias =
2439 layer->m_QuantizedLstmParameters.m_CellBias =
2440 std::make_shared<ScopedTensorHandle>(params.
GetCellBias());
2441 layer->m_QuantizedLstmParameters.m_OutputGateBias =
2451 const auto layer = m_Graph->AddLayer<
QLstmLayer>(descriptor, name);
2456 layer->m_BasicParameters.m_InputToCellWeights =
2458 layer->m_BasicParameters.m_InputToOutputWeights =
2460 layer->m_BasicParameters.m_RecurrentToForgetWeights =
2462 layer->m_BasicParameters.m_RecurrentToCellWeights =
2464 layer->m_BasicParameters.m_RecurrentToOutputWeights =
2466 layer->m_BasicParameters.m_ForgetGateBias =
2468 layer->m_BasicParameters.m_CellBias =
2469 std::make_shared<ScopedTensorHandle>(*(params.
m_CellBias));
2470 layer->m_BasicParameters.m_OutputGateBias =
2474 if(!descriptor.m_CifgEnabled)
2484 "AddQLstmLayer: Recurrent To Input Weights cannot be NULL");
2492 layer->m_CifgParameters.m_InputToInputWeights =
2494 layer->m_CifgParameters.m_RecurrentToInputWeights =
2496 layer->m_CifgParameters.m_InputGateBias =
2501 if(descriptor.m_ProjectionEnabled)
2508 layer->m_ProjectionParameters.m_ProjectionWeights =
2514 layer->m_ProjectionParameters.m_ProjectionBias =
2521 if(descriptor.m_PeepholeEnabled)
2533 if(!descriptor.m_CifgEnabled)
2540 layer->m_PeepholeParameters.m_CellToInputWeights =
2544 layer->m_PeepholeParameters.m_CellToForgetWeights =
2546 layer->m_PeepholeParameters.m_CellToOutputWeights =
2551 if(descriptor.m_LayerNormEnabled)
2568 if(!descriptor.m_CifgEnabled)
2575 layer->m_LayerNormParameters.m_InputLayerNormWeights =
2579 layer->m_LayerNormParameters.m_ForgetLayerNormWeights =
2581 layer->m_LayerNormParameters.m_CellLayerNormWeights =
2583 layer->m_LayerNormParameters.m_OutputLayerNormWeights =
2599 layer->Accept(visitor);
2607 layer->ExecuteStrategy(strategy);
2617 : m_Graph(
std::move(graph)), m_Guid(profiling::
ProfilingService::GetNextGuid()), m_ModelOptions(modelOptions)
A layer that the constant data can be bound to.
OptimizeForConnection< Layer, PermuteLayer, SquashEqualSiblingsImpl< PermuteLayer > > SquashEqualPermuteSiblings
void ReportError(const std::string &errorMessage, Optional< std::vector< std::string > &> errorMessages)
Iterator begin()
Returns iterator pointing to the beginning of the list. Lowercase for range-based for loops...
IConnectableLayer * AddSubtractionLayer(const char *name=nullptr)
Adds a subtraction layer to the network.
bool m_BiasEnabled
Enable/disable bias.
IConnectableLayer * AddReduceLayer(const ReduceDescriptor &reduceDescriptor, const char *name=nullptr)
ModelOptions m_ModelOptions
IConnectableLayer * AddRsqrtLayer(const char *name=nullptr)
IConnectableLayer * AddActivationLayer(const ActivationDescriptor &activationDescriptor, const char *name=nullptr)
Adds an activation layer to the network.
bool m_HalfPixelCenters
Half Pixel Centers.
bool m_AlignCorners
Aligned corners.
This layer represents a minimum operation.
static const FactoryId DeferredFactoryId
Use the workload factory to create the tensor handle.
This layer represents a split operation.
OptimizationResult AssignBackends(OptimizedNetworkImpl *optNetObjPtr, BackendSettings &backendSettings, Graph::Iterator &firstLayer, Graph::Iterator &lastLayer, Optional< std::vector< std::string > &> errMessages)
IConnectableLayer * AddCastLayer(const char *name=nullptr)
Adds a cast layer to the network.
LstmBasicParameters m_BasicParameters
const std::vector< InputSlot * > & GetConnections() const
FactoryFunction GetFactory(const BackendId &id) const
This layer represents a batch normalization operation.
static void Destroy(INetwork *network)
A ViewsDescriptor for the SplitterLayer.
Interface for a layer that is connectable to other layers via InputSlots and OutputSlots.
OptimizeForConnection< PermuteLayer, PermuteLayer, OptimizeInversePermutesImpl< PermuteLayer > > OptimizeInversePermutes
std::vector< ConvertFp32ToFp16Layer * > InsertConvertFp32ToFp16LayersAfter(Graph &graph, Layer &layer)
bool m_BiasEnabled
Enable/disable bias.
void ExecuteStrategy(IStrategy &strategy) const
void SetEdgeStrategy(unsigned int connectionIndex, EdgeStrategy strategy)
QuantizedLstmParameters m_QuantizedLstmParameters
This layer represents a 2D transpose convolution operation.
virtual Status PrintGraph()
No strategy has been defined. Used internally to verify integrity of optimizations.
std::vector< ConvertFp16ToFp32Layer * > InsertConvertFp16ToFp32LayersBefore(Graph &graph, Layer &layer, bool expectCorrectInputType)
IConnectableLayer * AddResizeLayer(const ResizeDescriptor &resizeDescriptor, const char *name=nullptr)
virtual ~OptimizedNetworkImpl()
A TransposeConvolution2dDescriptor for the TransposeConvolution2dLayer.
const TensorShape & GetShape() const
CPU Execution: Reference C++ kernels.
IConnectableLayer * AddTransposeConvolution2dLayer(const TransposeConvolution2dDescriptor &descriptor, const ConstTensor &weights, const Optional< ConstTensor > &biases, const char *name=nullptr)
Adds a 2D transpose convolution layer to the network.
IConnectableLayer * AddQuantizeLayer(const char *name=nullptr)
OptimizeForExclusiveConnection< PadLayer, Convolution2dLayer, pad_fold::FoldPadIntoConvolution2dImpl > FoldPadIntoConvolution2d
Optimizer::Optimizations MakeOptimizations(Args &&... args)
IConnectableLayer * AddQLstmLayer(const QLstmDescriptor &descriptor, const LstmInputParams ¶ms, const char *name=nullptr)
Add a QLstm layer to the network.
IConnectableLayer * AddRankLayer(const char *name=nullptr)
A ReshapeDescriptor for the ReshapeLayer.
OptimizeForConnection< TransposeLayer, TransposeLayer, OptimizeInversePermutesImpl< TransposeLayer > > OptimizeInverseTransposes
IConnectableLayer * AddAdditionLayer(const char *name=nullptr)
OptimizeForConnection< TransposeLayer, BatchToSpaceNdLayer, PermuteAndBatchToSpaceAsDepthToSpaceImpl< TransposeLayer > > TransposeAndBatchToSpaceAsDepthToSpace
OptimizeForExclusiveConnection< DepthwiseConvolution2dLayer, BatchNormalizationLayer, FuseBatchNorm< DepthwiseConvolution2dLayer, armnn::DataType::Float32 > > FuseBatchNormIntoDepthwiseConvolution2DFloat32
IConnectableLayer * AddConstantLayer(const ConstTensor &input, const char *name=nullptr)
Adds a layer with no inputs and a single output, which always corresponds to the passed in constant t...
IConnectableLayer * AddFullyConnectedLayer(const FullyConnectedDescriptor &fullyConnectedDescriptor, const Optional< ConstTensor > &weights, const Optional< ConstTensor > &biases, const char *name=nullptr)
Adds a fully connected layer to the network.
#define ARMNN_NO_DEPRECATE_WARN_BEGIN
IConnectableLayer * AddDepthwiseConvolution2dLayer(const DepthwiseConvolution2dDescriptor &convolution2dDescriptor, const ConstTensor &weights, const Optional< ConstTensor > &biases, const char *name=nullptr)
Adds a 2D depthwise convolution layer to the network.
ITensorHandleFactory::FactoryId CalculateSlotOptionForOutput(BackendsMap &backends, OutputSlot &slot, TensorHandleFactoryRegistry ®istry)
A ComparisonDescriptor for the ComparisonLayer.
IConnectableLayer * AddAbsLayer(const char *name=nullptr)
This layer represents a depthwise convolution 2d operation.
static void ConvertBFloat16ToFloat32(const void *srcBFloat16Buffer, size_t numElements, float *dstFloat32Buffer)
bool RequiresCopy(ITensorHandleFactory::FactoryId src, ITensorHandleFactory::FactoryId dst, TensorHandleFactoryRegistry ®istry)
IConnectableLayer * AddPooling2dLayer(const Pooling2dDescriptor &pooling2dDescriptor, const char *name=nullptr)
NetworkImpl(NetworkOptions networkOptions={})
IConnectableLayer * AddStackLayer(const StackDescriptor &descriptor, const char *name=nullptr)
Adds a stack layer to the network.
std::shared_ptr< ConstTensorHandle > m_LayerOutput
uint32_t m_TargetWidth
Target width value.
DataLayout m_DataLayout
The data layout to be used (NCHW, NHWC).
std::vector< BackendOptions > ModelOptions
void Accept(ILayerVisitor &visitor) const
IConnectableLayer * AddMergeLayer(const char *name=nullptr)
IConnectableLayer * AddRsqrtLayer(const char *name=nullptr)
Add Reciprocal of square root layer to the network.
A Convolution2dDescriptor for the Convolution2dLayer.
Layer & GetOwningLayer() const
Source backends tensor data can be exported to destination backend tensor without copy...
IConnectableLayer * AddQuantizeLayer(const char *name=nullptr)
Add a quantize layer to the network.
This layer converts data type Float 16 to Float 32.
IConnectableLayer * AddMinimumLayer(const char *name=nullptr)
Add a Minimum layer to the network.
IConnectableLayer * AddQuantizedLstmLayer(const QuantizedLstmInputParams ¶ms, const char *name=nullptr)
IConnectableLayer * AddL2NormalizationLayer(const L2NormalizationDescriptor &desc, const char *name=nullptr)
Adds an L2 normalization layer to the network.
bool m_BiasEnabled
Enable/disable bias.
IConnectableLayer * AddConstantLayer(const ConstTensor &input, const char *name=nullptr)
IConnectableLayer * AddElementwiseUnaryLayer(const ElementwiseUnaryDescriptor &elementwiseUnaryDescriptor, const char *name=nullptr)
IConnectableLayer * AddMergeLayer(const char *name=nullptr)
Adds a merge layer to the network.
IConnectableLayer * AddPermuteLayer(const PermuteDescriptor &permuteDescriptor, const char *name=nullptr)
Adds a permute layer to the network.
ResizeMethod m_Method
The Interpolation method to use (Bilinear, NearestNeighbor).
IConnectableLayer * AddNormalizationLayer(const NormalizationDescriptor &normalizationDescriptor, const char *name=nullptr)
IConnectableLayer * AddSliceLayer(const SliceDescriptor &sliceDescriptor, const char *name=nullptr)
Adds a slice layer to the network.
static void Pass(Graph &graph, const Optimizations &optimizations)
OptimizeForExclusiveConnection< DepthwiseConvolution2dLayer, BatchNormalizationLayer, FuseBatchNorm< DepthwiseConvolution2dLayer, armnn::DataType::Float16 > > FuseBatchNormIntoDepthwiseConvolution2DFloat16
IConnectableLayer * AddFloorLayer(const char *name=nullptr)
Adds a floor layer to the network.
IConnectableLayer * AddPooling2dLayer(const Pooling2dDescriptor &pooling2dDescriptor, const char *name=nullptr)
Adds a pooling layer to the network.
This layer represents a SpaceToDepth operation.
IConnectableLayer * AddMeanLayer(const MeanDescriptor &meanDescriptor, const char *name=nullptr)
This layer represents a reshape operation.
void ExecuteStrategy(IStrategy &strategy) const
OptimizeForExclusiveConnection< Convolution2dLayer, BatchNormalizationLayer, FuseBatchNorm< Convolution2dLayer, armnn::DataType::Float16 > > FuseBatchNormIntoConvolution2DFloat16
OptimizeForExclusiveConnection< Convolution2dLayer, BatchNormalizationLayer, FuseBatchNorm< Convolution2dLayer, armnn::DataType::Float32 > > FuseBatchNormIntoConvolution2DFloat32
#define ARMNN_LOG(severity)
IConnectableLayer * AddMinimumLayer(const char *name=nullptr)
std::shared_ptr< ConstTensorHandle > m_Weight
A unique pointer to store Weight values.
IConnectableLayer * AddFillLayer(const FillDescriptor &fillDescriptor, const char *name=nullptr)
Add an Fill layer to the network.
Main network class which provides the interface for building up a neural network. ...
This layer represents an activation operation with the specified activation function.
IConnectableLayer * AddSpaceToDepthLayer(const SpaceToDepthDescriptor &spaceToDepthDescriptor, const char *name=nullptr)
BackendRegistry & BackendRegistryInstance()
This layer converts data type BFloat16 to Float32.
LayerT * ConvertBf16ToFp32Weight(Layer *l)
std::vector< BackendOptions > NetworkOptions
std::shared_ptr< ConstTensorHandle > m_Mean
A unique pointer to store Mean values.
A LogicalBinaryDescriptor for the LogicalBinaryLayer.
IConnectableLayer * AddConvolution2dLayer(const Convolution2dDescriptor &convolution2dDescriptor, const ConstTensor &weights, const Optional< ConstTensor > &biases, const char *name=nullptr)
Adds a 2D convolution layer to the network.
This layer represents an unknown operation in the input graph.
OptimizeForConnection< Layer, ReshapeLayer, SquashEqualSiblingsImpl< ReshapeLayer > > SquashEqualReshapeSiblings
IConnectableLayer * AddLogSoftmaxLayer(const LogSoftmaxDescriptor &logSoftmaxDescriptor, const char *name=nullptr)
Adds a log softmax layer to the network.
IConnectableLayer * AddResizeLayer(const ResizeDescriptor &resizeDescriptor, const char *name=nullptr)
Adds a resize layer to the network.
IConnectableLayer * AddActivationLayer(const ActivationDescriptor &activationDescriptor, const char *name=nullptr)
This layer represents a detection postprocess operator.
BackendIdSet m_SupportedBackends
OptimizeForConnection< Layer, TransposeLayer, MoveTransposeUpImpl > MoveTransposeUp
IConnectableLayer * AddBatchToSpaceNdLayer(const BatchToSpaceNdDescriptor &batchToSpaceNdDescriptor, const char *name=nullptr)
Adds a batch to space ND layer to the network.
OptimizationResult ReturnWithError(OptimizationResult res, const Layer *layer, const BackendSettings &backendSettings, Optional< std::vector< std::string > &> errMessages)
Copyright (c) 2021 ARM Limited and Contributors.
This layer represents a pad operation.
This layer represents a LSTM operation.
void IgnoreUnused(Ts &&...)
IConnectableLayer * AddDivisionLayer(const char *name=nullptr)
Adds a division layer to the network.
void SetBackendId(const BackendId &id)
const std::vector< InputSlot > & GetInputSlots() const
bool IsBackendSupported(const BackendId &backend) const
LayerList::const_iterator Iterator
This layer represents a reduction operation.
IConnectableLayer * AddDepthwiseConvolution2dLayer(const DepthwiseConvolution2dDescriptor &convolution2dDescriptor, const ConstTensor &weights, const Optional< ConstTensor > &biases, const char *name=nullptr)
IConnectableLayer * AddLogSoftmaxLayer(const LogSoftmaxDescriptor &logSoftmaxDescriptor, const char *name=nullptr)
A SpaceToDepthDescriptor for the SpaceToDepthLayer.
This layer represents a permutation operation.
unsigned int GetNumOutputSlots() const override
Returns the number of connectable output slots.
This layer represents a SpaceToBatchNd operation.
A BatchToSpaceNdDescriptor for the BatchToSpaceNdLayer.
OptimizeForType< Layer, AddDebugImpl > InsertDebugLayer
virtual Status SerializeToDot(std::ostream &stream) const
OptimizeForConnection< ReshapeLayer, ReshapeLayer, OptimizeConsecutiveReshapesImpl > OptimizeConsecutiveReshapes
IConnectableLayer * AddMeanLayer(const MeanDescriptor &meanDescriptor, const char *name=nullptr)
Add a Mean layer to the network.
Private implementation of INetwork.
int LayerBindingId
Type of identifiers for bindable layers (inputs, outputs).
IConnectableLayer * AddInputLayer(LayerBindingId id, const char *name=nullptr)
Adds an input layer to the network.
This layer represents a elementwiseUnary operation.
constexpr const char * GetDataTypeName(DataType dataType)
IConnectableLayer * AddStridedSliceLayer(const StridedSliceDescriptor &stridedSliceDescriptor, const char *name=nullptr)
Adds a strided slice layer to the network.
A ResizeDescriptor for the ResizeLayer.
A StackDescriptor for the StackLayer.
Destination backend can work directly with tensors on source backend.
virtual std::vector< ITensorHandleFactory::FactoryId > GetHandleFactoryPreferences() const
(Optional) Returns a vector of supported TensorHandleFactory ids in preference order.
IConnectableLayer * AddSpaceToDepthLayer(const SpaceToDepthDescriptor &spaceToDepthDescriptor, const char *name=nullptr)
Adds a space to depth layer to the network.
OptimizeForConnection< ConvertFp16ToFp32Layer, ConvertFp32ToFp16Layer, OptimizeInverseConversionsImpl > OptimizeInverseConversionsFp16
IConnectableLayer * AddSoftmaxLayer(const SoftmaxDescriptor &softmaxDescriptor, const char *name=nullptr)
Adds a softmax layer to the network.
The SubgraphView class represents a subgraph of a Graph.
IConnectableLayer * AddMergerLayer(const MergerDescriptor &mergerDescriptor, const char *name=nullptr)
profiling::ProfilingGuid GetGuid() const
IConnectableLayer * AddFloorLayer(const char *name=nullptr)
A PadDescriptor for the PadLayer.
This layer represents an instance normalization operation.
std::shared_ptr< ConstTensorHandle > m_InputToForgetWeights
A unique pointer to represent 2D weights tensor with dimensions [num_units, inputSize] (QSymmS8)...
IConnectableLayer * AddMultiplicationLayer(const char *name=nullptr)
OptimizeForConnection< PermuteLayer, BatchToSpaceNdLayer, PermuteAndBatchToSpaceAsDepthToSpaceImpl< PermuteLayer > > PermuteAndBatchToSpaceAsDepthToSpace
IConnectableLayer * AddQLstmLayer(const QLstmDescriptor &descriptor, const LstmInputParams ¶ms, const char *name=nullptr)
OptimizeForConnection< Layer, PermuteLayer, MovePermuteUpImpl > MovePermuteUp
This layer represents a Logical Binary operation.
IConnectableLayer * AddStridedSliceLayer(const StridedSliceDescriptor &stridedSliceDescriptor, const char *name=nullptr)
std::unique_ptr< NetworkImpl > pNetworkImpl
A layer user-provided data can be bound to (e.g. inputs, outputs).
IOptimizedNetwork(std::unique_ptr< Graph > graph)
IConnectableLayer * AddResizeBilinearLayer(const ResizeBilinearDescriptor &resizeDesc, const char *name=nullptr)
Adds a resize bilinear layer to the network.
friend IOptimizedNetworkPtr Optimize(const INetwork &inNetwork, const std::vector< BackendId > &backendPreferences, const IDeviceSpec &deviceSpec, const OptimizerOptions &options, Optional< std::vector< std::string > &> messages)
Create an optimized version of the network.
void ForEachLayer(Func func) const
virtual std::vector< Capability > GetCapabilities(const IConnectableLayer *layer, const IConnectableLayer *connectedLayer, CapabilityClass capabilityClass)
IConnectableLayer * AddInputLayer(LayerBindingId id, const char *name=nullptr)
This layer dequantizes the input tensor.
ConvertConstants< Float32ToFloat16, IsFloat16Layer > ConvertConstantsFloatToHalf
OptimizeForType< TransposeLayer, TransposeAsReshapeImpl > TransposeAsReshape
This layer represents a Gather operator.
This layer represents a fully connected operation.
An LstmDescriptor for the LstmLayer.
#define ARMNN_NO_DEPRECATE_WARN_END
std::shared_ptr< ConstTensorHandle > m_Weight
A unique pointer to store Weight values.
IConnectableLayer * AddLstmLayer(const LstmDescriptor &descriptor, const LstmInputParams ¶ms, const char *name=nullptr)
Add a Lstm layer to the network.
IConnectableLayer * AddTransposeConvolution2dLayer(const TransposeConvolution2dDescriptor &descriptor, const ConstTensor &weights, const Optional< ConstTensor > &biases, const char *name=nullptr)
#define ARMNN_ASSERT_MSG(COND, MSG)
IConnectableLayer * AddRankLayer(const char *name=nullptr)
Adds a rank layer to the network.
This layer represents a QuantizedLstm operation.
This layer represents a log softmax operation.
OptimizationResult ApplyBackendOptimizations(OptimizedNetworkImpl *optNetObjPtr, BackendSettings &backendSettings, BackendsMap &backends, const ModelOptions &modelOptions, Optional< std::vector< std::string > &> errMessages)
IConnectableLayer * AddPreluLayer(const char *name=nullptr)
A L2NormalizationDescriptor for the L2NormalizationLayer.
int32_t GetQuantizationOffset() const
An ArgMinMaxDescriptor for ArgMinMaxLayer.
float GetQuantizationScale() const
DataType GetDataType() const
An OriginsDescriptor for the ConcatLayer.
A ReduceDescriptor for the REDUCE operators.
IConnectableLayer * AddDetectionPostProcessLayer(const DetectionPostProcessDescriptor &descriptor, const ConstTensor &anchors, const char *name=nullptr)
bool has_value() const noexcept
A FullyConnectedDescriptor for the FullyConnectedLayer.
bool m_BiasEnabled
Enable/disable bias.
This layer represents a stack operation.
IConnectableLayer * AddFullyConnectedLayer(const FullyConnectedDescriptor &fullyConnectedDescriptor, const Optional< ConstTensor > &weights, const Optional< ConstTensor > &biases, const char *name=nullptr)
A tensor defined by a TensorInfo (shape and data type) and an immutable backing store.
const Subgraphs & GetFailedSubgraphs() const
This layer represents a merge operation.
This layer represents a softmax operation.
IConnectableLayer * AddInstanceNormalizationLayer(const InstanceNormalizationDescriptor &desc, const char *name=nullptr)
Adds an instance normalization layer to the network.
const std::string & GetNameStr() const
LayerType GetType() const override
Returns the armnn::LayerType of this layer.
uint32_t m_TargetWidth
Target width value.
std::vector< ConvertBf16ToFp32Layer * > InsertConvertBf16ToFp32LayersBefore(Graph &graph, Layer &layer, bool expectCorrectInputType)
A GatherDescriptor for the GatherLayer.
std::unique_ptr< OptimizedNetworkImpl > pOptimizedNetworkImpl
IConnectableLayer * AddDivisionLayer(const char *name=nullptr)
This layer represents a BatchToSpaceNd operation.
IConnectableLayer * AddOutputLayer(LayerBindingId id, const char *name=nullptr)
Adds an output layer to the network.
std::vector< SubgraphViewPtr > Subgraphs
std::unique_ptr< IOptimizedNetwork, void(*)(IOptimizedNetwork *network)> IOptimizedNetworkPtr
bool m_HalfPixelCenters
Half Pixel Centers.
IConnectableLayer * AddArgMinMaxLayer(const ArgMinMaxDescriptor &desc, const char *name=nullptr)
IConnectableLayer * AddOutputLayer(LayerBindingId id, const char *name=nullptr)
IConnectableLayer * AddDequantizeLayer(const char *name=nullptr)
void SetQuantizationScale(float scale)
This layer represents a ArgMinMax operation.
IConnectableLayer * AddConcatLayer(const ConcatDescriptor &concatDescriptor, const char *name=nullptr)
#define ARMNN_ASSERT(COND)
A StandInDescriptor for the StandIn layer.
A QLstmDescriptor for the QLstmLayer.
BackendIdVector GetAvailablePreferredBackends() const
Device specific knowledge to be passed to the optimizer.
IConnectableLayer * AddSwitchLayer(const char *name=nullptr)
static bool IsLayerSupported(const BackendId &backendId, const IConnectableLayer &layer, Optional< DataType > dataType, std::string &outReasonIfUnsupported)
IConnectableLayer * AddEqualLayer(const char *name=nullptr)
Add a Equal layer to the network.
IConnectableLayer * AddAbsLayer(const char *name=nullptr)
Add absolute layer to the network.
IConnectableLayer * AddSubtractionLayer(const char *name=nullptr)
IConnectableLayer * AddResizeBilinearLayer(const ResizeBilinearDescriptor &resizeDesc, const char *name=nullptr)
bool Validate(const SubgraphView &originalSubgraph) const
An ActivationDescriptor for the ActivationLayer.
IConnectableLayer * AddLogicalBinaryLayer(const LogicalBinaryDescriptor &logicalBinaryDescriptor, const char *name=nullptr)
std::vector< ConvertFp32ToBf16Layer * > InsertConvertFp32ToBf16LayersAfter(Graph &graph, Layer &layer)
const BackendId & GetBackendId() const
uint32_t m_TargetHeight
Target height value.
This layer represents a floor operation.
void Accept(ILayerVisitor &visitor) const
IConnectableLayer * AddBatchNormalizationLayer(const BatchNormalizationDescriptor &desc, const ConstTensor &mean, const ConstTensor &variance, const ConstTensor &beta, const ConstTensor &gamma, const char *name=nullptr)
Adds a batch normalization layer to the network.
INetwork(NetworkOptions networkOptions={})
uint32_t m_TargetHeight
Target height value.
A SliceDescriptor for the SliceLayer.
IConnectableLayer * AddSpaceToBatchNdLayer(const SpaceToBatchNdDescriptor &spaceToBatchNdDescriptor, const char *name=nullptr)
Adds a space to batch layer to the network.
This layer represents a normalization operation.
virtual MemorySourceFlags GetExportFlags() const
IConnectableLayer * AddStandInLayer(const StandInDescriptor &descriptor, const char *name=nullptr)
This layer represents a pooling 2d operation.
This layer converts data type Float 32 to Float 16.
This layer represents a transpose operation.
IConnectableLayer * AddPermuteLayer(const PermuteDescriptor &permuteDescriptor, const char *name=nullptr)
This layer represents an addition operation.
IConnectableLayer * AddComparisonLayer(const ComparisonDescriptor &comparisonDescriptor, const char *name=nullptr)
Add a Comparison layer to the network.
QLstmBasicParameters m_BasicParameters
void SubstituteSubgraph(SubgraphView &subgraph, IConnectableLayer *substituteLayer)
Substitutes the given sub-graph with either a new layer or a new sub-graph.
IConnectableLayer * AddL2NormalizationLayer(const L2NormalizationDescriptor &desc, const char *name=nullptr)
IConnectableLayer * AddArgMinMaxLayer(const ArgMinMaxDescriptor &desc, const char *name=nullptr)
Adds an ArgMinMax layer to the network.
bool CheckScaleSetOnQuantizedType(Layer *layer, Optional< std::vector< std::string > &> errMessages)
IConnectableLayer * AddSplitterLayer(const ViewsDescriptor &splitterDescriptor, const char *name=nullptr)
Adds a splitter layer to the network.
void SetTensorHandleFactory(const ITensorHandleFactory::FactoryId &id)
A SpaceToBatchNdDescriptor for the SpaceToBatchNdLayer.
OptimizeForType< PermuteLayer, PermuteAsReshapeImpl > PermuteAsReshape
bool IsWarningOnly() const
OptimizeForConnection< Layer, TransposeLayer, SquashEqualSiblingsImpl< TransposeLayer > > SquashEqualTransposeSiblings
This layer represents a QLstm operation.
IConnectableLayer * AddAdditionLayer(const char *name=nullptr)
Adds an addition layer to the network.
const Substitutions & GetSubstitutions() const
BackendIdVector m_PreferredBackends
This layer represents a subtraction operation.
IConnectableLayer * AddTransposeLayer(const TransposeDescriptor &transposeDescriptor, const char *name=nullptr)
This layer calculates both true and false outputs for input.
IConnectableLayer * AddReshapeLayer(const ReshapeDescriptor &reshapeDescriptor, const char *name=nullptr)
Adds a reshape layer to the network.
EmptyOptional is used to initialize the Optional class in case we want to have default value for an O...
IConnectableLayer * AddPadLayer(const PadDescriptor &padDescriptor, const char *name=nullptr)
Adds a fully pad layer to the network.
ConvertConstants< Float16ToFloat32, IsFloat32Layer > ConvertConstantsHalfToFloat
bool m_AlignCorners
Aligned corners.
A ElementwiseUnaryDescriptor for the ElementwiseUnaryLayer.
std::shared_ptr< ConstTensorHandle > m_InputToForgetWeights
A unique pointer to represent 2D weights tensor with dimensions [input_size, num_units].
static Subgraphs SelectSubgraphs(Graph &graph, const LayerSelectorFunction &selector)
Selects subgraphs from a graph based on the selector function and the algorithm.
IConnectableLayer * AddReshapeLayer(const ReshapeDescriptor &reshapeDescriptor, const char *name=nullptr)
This layer represents a L2 normalization operation.
This layer represents a cast operation.
IConnectableLayer * AddMaximumLayer(const char *name=nullptr)
std::shared_ptr< ConstTensorHandle > m_Weight
A unique pointer to store Weight values.
BackendsMap CreateSupportedBackends(TensorHandleFactoryRegistry &handleFactoryRegistry, BackendSettings &backendSettings)
IConnectableLayer * AddCastLayer(const char *name=nullptr)
OptimizeForConnection< ConvertFp32ToFp16Layer, ConvertFp16ToFp32Layer, OptimizeInverseConversionsImpl > OptimizeInverseConversionsFp32
IConnectableLayer * AddStackLayer(const StackDescriptor &stackDescriptor, const char *name=nullptr)
IConnectableLayer * AddSoftmaxLayer(const SoftmaxDescriptor &softmaxDescriptor, const char *name=nullptr)
const std::string & Get() const
IConnectableLayer * AddGreaterLayer(const char *name=nullptr)
IConnectableLayer * AddNormalizationLayer(const NormalizationDescriptor &normalizationDescriptor, const char *name=nullptr)
Adds a normalization layer to the network.
BackendIdSet m_SelectedBackends
OptimizeForExclusiveConnection< PadLayer, Pooling2dLayer, pad_fold::FoldPadIntoPooling2dImpl > FoldPadIntoPooling2d
IConnectableLayer * AddEqualLayer(const char *name=nullptr)
IConnectableLayer * AddFillLayer(const FillDescriptor &fillDescriptor, const char *name=nullptr)
Iterator end()
Returns iterator pointing to the end of the list. Lowercase for range-based for loops.
const Graph & GetGraph() const
OptimizationResult AttemptBackendAssignment(BackendSettings &backendSettings, Graph &graph, Layer *layer, BackendId backend, DataType dataTypeIn, DataType dataTypeOut, const std::vector< BackendId > &availablePreferredBackends, std::string &reasonIfUnsupported, Optional< std::vector< std::string > &> errMessages)
ITensorHandleFactory * GetFactory(ITensorHandleFactory::FactoryId id) const
Find a TensorHandleFactory by Id Returns nullptr if not found.
void SetTensorInfo(const TensorInfo &tensorInfo) override
OptimizedNetworkImpl(std::unique_ptr< Graph > graph)
A MeanDescriptor for the MeanLayer.
This layer represents a division operation.
Status SerializeToDot(std::ostream &stream) const
IConnectableLayer * AddSpaceToBatchNdLayer(const SpaceToBatchNdDescriptor &spaceToBatchNdDescriptor, const char *name=nullptr)
IConnectableLayer * AddQuantizedLstmLayer(const QuantizedLstmInputParams ¶ms, const char *name=nullptr)
Add a QuantizedLstm layer to the network.
This layer represents a strided slice operation.
ITensorHandleFactory::FactoryId CalculateSlotOption(BackendsMap &backends, OutputSlot &outputSlot, TensorHandleFactoryRegistry ®istry, bool importEnabled)
This layer represents a maximum operation.
const OutputSlot & GetOutputSlot(unsigned int index=0) const override
Get the const output slot handle by slot index.
OptimizeForType< Layer, ConvertFp32NetworkToFp16Impl > Fp32NetworkToFp16Converter
A TransposeDescriptor for the TransposeLayer.
A StridedSliceDescriptor for the StridedSliceLayer.
IConnectableLayer * AddComparisonLayer(const ComparisonDescriptor &comparisonDescriptor, const char *name=nullptr)
OptimizationResult SelectTensorHandleStrategy(Graph &optGraph, BackendsMap &backends, TensorHandleFactoryRegistry ®istry, bool importEnabled, Optional< std::vector< std::string > &> errMessages)
void ReportWarning(const std::string &warningMessage, Optional< std::vector< std::string > &> warningMessages)
This layer represents a convolution 2d operation.
This layer converts data type Float32 to BFloat16.
void SetQuantizationOffset(int32_t offset)
IConnectableLayer * AddSwitchLayer(const char *name=nullptr)
Adds a switch layer to the network.
IConnectableLayer * AddDetectionPostProcessLayer(const DetectionPostProcessDescriptor &descriptor, const ConstTensor &anchors, const char *name=nullptr)
Adds a Detection PostProcess layer to the network.
static INetwork * CreateRaw(NetworkOptions networkOptions={})
IConnectableLayer * AddMergerLayer(const MergerDescriptor &mergerDescriptor, const char *name=nullptr)
Adds a concat layer to the network.
IConnectableLayer * AddMultiplicationLayer(const char *name=nullptr)
Adds a multiplication layer to the network.
IConnectableLayer * AddStandInLayer(const StandInDescriptor &descriptor, const char *name=nullptr)
Add a stand-in layer for a type unknown to the Arm NN framework.
This layer represents a mean operation.
This layer represents a comparison operation.
std::unique_ptr< INetwork, void(*)(INetwork *network)> INetworkPtr
IConnectableLayer * AddBatchNormalizationLayer(const BatchNormalizationDescriptor &desc, const ConstTensor &mean, const ConstTensor &variance, const ConstTensor &beta, const ConstTensor &gamma, const char *name=nullptr)
IConnectableLayer * AddReduceLayer(const ReduceDescriptor &reduceDescriptor, const char *name=nullptr)
Adds a reduce layer to the network.
IConnectableLayer * AddElementwiseUnaryLayer(const ElementwiseUnaryDescriptor &elementwiseUnaryDescriptor, const char *name=nullptr)
Add an ElementwiseUnary layer to the network.
OptimizeForType< Layer, AddBroadcastReshapeLayerImpl > AddBroadcastReshapeLayer
IConnectableLayer * AddGatherLayer(const char *name=nullptr)
IConnectableLayer * AddDequantizeLayer(const char *name=nullptr)
Adds a Dequantize layer to the network.
A Pooling2dDescriptor for the Pooling2dLayer.
This layer dequantizes the input tensor.
A NormalizationDescriptor for the NormalizationLayer.
IConnectableLayer * AddPreluLayer(const char *name=nullptr)
Adds a PReLU layer to the network.
IConnectableLayer * AddGreaterLayer(const char *name=nullptr)
Add a Greater layer to the network.
IConnectableLayer * AddDepthToSpaceLayer(const DepthToSpaceDescriptor &depthToSpaceDescriptor, const char *name=nullptr)
IConnectableLayer * AddMaximumLayer(const char *name=nullptr)
Add a Maximum layer to the network.
DataLayout m_DataLayout
The data layout to be used (NCHW, NHWC).
An InstanceNormalizationDescriptor for InstanceNormalizationLayer.
This layer represents a multiplication operation.
IConnectableLayer * AddConvolution2dLayer(const Convolution2dDescriptor &convolution2dDescriptor, const ConstTensor &weights, const Optional< ConstTensor > &biases, const char *name=nullptr)
std::shared_ptr< ConstTensorHandle > m_Anchors
A unique pointer to store Anchor values.
IConnectableLayer * AddSplitterLayer(const ViewsDescriptor &splitterDescriptor, const char *name=nullptr)
IConnectableLayer * AddBatchToSpaceNdLayer(const BatchToSpaceNdDescriptor &batchToSpaceNdDescriptor, const char *name=nullptr)
A ResizeBilinearDescriptor for the ResizeBilinearLayer.
IConnectableLayer * AddLstmLayer(const LstmDescriptor &descriptor, const LstmInputParams ¶ms, const char *name=nullptr)
IConnectableLayer * AddPadLayer(const PadDescriptor &padDescriptor, const char *name=nullptr)
std::shared_ptr< ConstTensorHandle > m_InputToInputWeights
A unique pointer to represent 2D weights tensor with dimensions [outputSize, inputSize] (QAsymm8)...
const TensorInfo & GetTensorInfo() const override
static INetworkPtr Create(NetworkOptions networkOptions={})
IConnectableLayer * AddLogicalBinaryLayer(const LogicalBinaryDescriptor &descriptor, const char *name=nullptr)
Adds a Logical Binary layer to the network.
IConnectableLayer * AddGatherLayer(const char *name=nullptr)
Add Gather layer to the network.
EdgeStrategy CalculateEdgeStrategy(BackendsMap &backends, ITensorHandleFactory::FactoryId srcFactoryId, const Layer &layer, const Layer &connectedLayer, TensorHandleFactoryRegistry ®istry, bool importEnabled)
static void Destroy(IOptimizedNetwork *network)
IConnectableLayer * AddConcatLayer(const ConcatDescriptor &concatDescriptor, const char *name=nullptr)
Adds a concatenation layer to the network.
virtual MemorySourceFlags GetImportFlags() const
OptimizeForType< Layer, ConvertFp32NetworkToBf16Impl > Fp32NetworkToBf16Converter
A SoftmaxDescriptor for the SoftmaxLayer.
const char * GetLayerTypeAsCString(LayerType type)
virtual bool SupportsMapUnmap() const
void AddCompatibilityLayers(std::map< BackendId, std::unique_ptr< class IBackendInternal >> &backends, TensorHandleFactoryRegistry ®istry)
Modifies the graph in-place, removing edges connecting layers using different compute devices...
std::shared_ptr< ConstTensorHandle > m_Weight
A unique pointer to store weight values.
bool IsCpuRefUsed() const
static const FactoryId LegacyFactoryId
ITensorHandleFactory::FactoryId CalculateSlotOptionForInput(BackendsMap &backends, OutputSlot &slot, TensorHandleFactoryRegistry ®istry, bool importEnabled)
This layer represents a fill operation.
A DepthwiseConvolution2dDescriptor for the DepthwiseConvolution2dLayer.
A FillDescriptor for the FillLayer.
This layer represents a DepthToSpace operation.
A BatchNormalizationDescriptor for the BatchNormalizationLayer.
IConnectableLayer * AddTransposeLayer(const TransposeDescriptor &transposeDescriptor, const char *name=nullptr)
Adds a transpose layer to the network.
unsigned int GetNumElements() const
std::map< BackendId, std::unique_ptr< class IBackendInternal > > BackendsMap
This layer represents a resize operation.
A PermuteDescriptor for the PermuteLayer.
IConnectableLayer * AddSliceLayer(const SliceDescriptor &sliceDescriptor, const char *name=nullptr)
IConnectableLayer * AddDepthToSpaceLayer(const DepthToSpaceDescriptor &depthToSpaceDescriptor, const char *name=nullptr)
Adds a depth to space layer to the network.
LayerType
When adding a new layer, adapt also the LastLayer enum value in the enum class LayerType below...
bool m_ConstantWeights
Enable/disable constant weights and biases.
IConnectableLayer * AddInstanceNormalizationLayer(const InstanceNormalizationDescriptor &desc, const char *name=nullptr)
std::vector< float > anchors({ 0.5f, 0.5f, 1.0f, 1.0f, 0.5f, 0.5f, 1.0f, 1.0f, 0.5f, 0.5f, 1.0f, 1.0f, 0.5f, 10.5f, 1.0f, 1.0f, 0.5f, 10.5f, 1.0f, 1.0f, 0.5f, 100.5f, 1.0f, 1.0f })